Jul 22, 2026

Daily Updates

• Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber
  Gemini 3.6 Flash enhances efficiency and quality, achieving 17% fewer output tokens than its predecessor, 3.5 Flash, while also reducing costs to $1.50/1M input tokens and $7.50/1M output tokens.

• Kimi K3 Is Competitive with Fable; Kimi K3 and Fable Is SoTA
  Kimi K3 competes effectively with Fable, achieving 93% accuracy on over 1,000 tasks, while their combination creates a state-of-the-art (SoTA) solution that outperforms either model alone.

• Judge approves $1.5B Anthropic settlement for pirated books used to train Claude
  A federal judge has approved a $1.5 billion settlement requiring Anthropic to compensate authors approximately $3,000 per book for using pirated works to train its Claude chatbot, marking a significant legal precedent in AI copyright cases.

• Laguna S 2.1
  Laguna S 2.1 is a 118B parameter Mixture-of-Experts model that excels in long-horizon tasks, achieving a 70.2% score on Terminal-Bench 2.1, outperforming larger models in its class.

• GigaToken: ~1000x faster Language model tokenization
  Gigatoken is a drop-in replacement tokenizer that achieves speeds of ~1000x faster than HuggingFace's tokenizers, enabling text data processing at GB/s rates across various CPU architectures.

• Gemini last models: temperature, top_p, and top_k are deprecated and ignored
  Gemini 3.6 Flash and 3.5 Flash-Lite models are now available, offering enhanced performance for complex tasks and lower costs, with 3.6 Flash priced at $1.50/1M input tokens and $7.50/1M output tokens.

• Can a MUD evaluate LLMs? A $99 proof of concept
  CrucibleBench introduces a benchmark for AI behavior in a persistent MUD, where NPCs remember actions and trust accumulates, allowing for nuanced evaluation over 50 turns with hidden social objectives.

• Kimi K3: second only to Fable 5 on AA-Briefcase
  Kimi K3, with 2.8T parameters, scores 1543 on AA-Briefcase, outperforming models like GPT-5.6 and Opus 4.8, but incurs a high operational cost of $10.57 per task.

• I Built a Private Genomics Study with Stoffel MPC
  Stoffel MPC enables genomic studies without collecting raw DNA, allowing participants to contribute data securely through a decentralized model that maintains privacy while computing aggregate statistics.

• Built for Vera Rubin, NVIDIA Spectrum-6 Arrives in Gigascale AI Factories
  NVIDIA Spectrum-6 is a 102.4-terabit-per-second Ethernet switch that doubles the capacity of its predecessor, designed to enhance the performance and efficiency of gigascale AI factories by connecting vast numbers of GPUs and CPUs.

• Atomically Thin Materials Significantly Shrink Qubits
  MIT researchers have demonstrated that using 2D materials like hexagonal boron nitride in capacitors can significantly reduce the size of qubits, enhancing the scalability of quantum processors.

• Introducing Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber
  Gemini 3.6 Flash enhances efficiency by reducing output token usage by 17% compared to its predecessor, 3.5 Flash, while also lowering costs to $1.50/1M input tokens and $7.50/1M output tokens.

• NVIDIA Vera Rubin Driving Performance Per Watt, Lowest Token Cost for Partners Worldwide
  NVIDIA's Vera Rubin platform achieves benchmark leadership in performance per watt and offers the lowest token costs for partners like CoreWeave and Google Cloud, enhancing AI infrastructure efficiency globally.

• NVIDIA Open Sources First GPU-Accelerated Medical Physics Simulation Framework
  NVIDIA's new Medical Physics Simulation framework enables developers to create reusable simulation environments for healthcare robotics, significantly enhancing the training process for surgical robots by simulating complex anatomical interactions and device behaviors.

• The State of Simulation for Physical AI: An Overview
  Simulation is essential for training physical AI systems, as it allows for the generation of vast amounts of photorealistic data, overcoming the limitations of real-world data collection which is often slow and costly.